Deep Learning Computed Tomography: Learning Projection-Domain Weights From Image Domain in Limited Angle Problems.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• 3-D tomographic reconstruction is crucial for medical imaging.
• Existing deep learning methods face challenges with memory requirements for back-projection.
• Analytical algorithms often struggle with limited-angle reconstruction.

Purpose of the Study:

• To develop a novel deep learning framework for 3-D tomographic reconstruction.
• To address memory limitations in back-projection layers.
• To enable joint optimization in both volume and projection domains.

Main Methods:

• Mapping filtered back-projection algorithms to neural networks.
• Introducing a new cone-beam back-projection layer for efficient forward and backward passes.
• Utilizing a public dataset for numerical evaluation in a limited-angle setting.

Main Results:

• Consistent improvement over analytical algorithms in limited-angle reconstruction.
• Achieved a 23% increase in peak signal-to-noise ratio in the region of interest.
• Demonstrated that the network learns interpretable strategies like compensation weights and apodization.

Conclusions:

• The proposed deep learning framework offers superior performance for 3-D tomographic reconstruction.
• The novel back-projection layer overcomes memory constraints and enables joint optimization.
• The framework provides interpretable results, aligning with established cone-beam reconstruction concepts.